Hardware

Nvidia CMP 170HX modified mining GPU beats A100 in FP16 benchmark

By One Percent LowsCovered by 2 outlets
★ Add us as a preferred source on Google

TL;DR

Hardware researcher Kate used voltage fault injection to unlock 80 GB of HBM and more processing units on Nvidia's CMP 170HX mining card, beating a retail A100 in an FP16 test.

Hardware researcher Kate has modified Nvidia's CMP 170HX mining graphics card to outperform a retail Nvidia A100 accelerator in a specific benchmark. The modified card achieved 85.9 TFLOPS in a ProjectPhysX OpenCL FP16 test, compared to approximately 72.5 TFLOPS for the A100. This is an 18% performance advantage in the tested workload.

The achievement relies on a voltage fault injection technique to bypass hardware restrictions built into the GA100 graphics processor. Out of ten tested 8GB cards, nine successfully exposed 80GB of HBM memory, while one reached 64GB due to a defective physical memory component. Earlier software tools like CMPUnlocker previously allowed some 8GB cards to expose 64GB and 10GB models to access 40GB. The new hardware modification also exposed approximately 120 to 126 working streaming multiprocessors.

Reports say secondary market prices for the CMP 170HX climbed from about $100 to over $1,000 after unlocking tools appeared. Despite the benchmark win, the modified card is not a complete replacement for an A100. Features like Error Correcting Code memory, Multi-Instance GPU support, video processing, PCIe, and NVLink functionality require additional work and remain untested.

More on NVIDIA